rosotacom-dev 2.5.dev75__tar.gz → 2.5.dev77__tar.gz

{rosotacom_dev-2.5.dev75 → rosotacom_dev-2.5.dev77}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: rosotacom-dev
-Version: 2.5.dev75
+Version: 2.5.dev77
 Summary: Host CLI for ROS OTA communication sessions.
 Author: develNor
 License-Expression: BSD-3-Clause

{rosotacom_dev-2.5.dev75 → rosotacom_dev-2.5.dev77}/docs/findings/README.md

@@ -28,6 +28,9 @@ living verification hook.
 
 ## Findings
 
+- [The size law belongs to the transport, not to the link](the-size-law-belongs-to-the-transport-not-the-link.md) - one emulated link, three OTA middlewares: CycloneDDS follows `1-(1-p)^N` with a flat residual over three decades of size, Fast DDS keeps the form and loses the fit, and zenoh is flat from 300 B to 12 kB and then falls off a cliff at 38 kB.
+- [Neither a queue nor a burst tilts the size law](neither-a-queue-nor-a-burst-tilts-the-size-law.md) - the field's size-systematic residual is reproduced by none of the three candidates: independent per-packet loss fits flat, a token bucket a burst can fill changes nothing systematic, and a serialisation constraint does not tilt the curve but truncates it (38 kB goes from 32% to 97.5% while 1 kB pays 10% more).
+- [A shorter group does not lose fewer key frames](a-shorter-group-does-not-lose-fewer-key-frames.md) - shortening the group cuts the large message's loss from 29.5% to 12.5% and leaves the rate at which large messages go missing unchanged (0.59, 0.79, 0.62 per second), while doubling the link's per-packet rate moves it by 1.8x.
 - [A fitted delay table never reached the shaper](a-fitted-delay-table-never-reached-the-shaper.md) - every profile carrying a custom netem distribution failed at arming: the table was installed as `/usr/lib/tc/<name>.dist` while iproute2 on a multiarch image reads `/usr/lib/<triplet>/tc`, so the drive-fitted delay tail was never actually in use.
 - [Sender-side transit rows made a dead link report zero loss](sender-rows-make-a-dead-link-look-loss-free.md) - `delivered = rows - lost` counted the sender's own `sent` rows as deliveries, so a topic the receiver never saw summarised as `delivered == expected` at 0.0% loss, with the publisher's timer reported as arrival spacing.
 - [A bridge that learns types from the graph cannot use a transport that does not carry the graph](graph-derived-types-block-a-non-graph-transport.md) - over `zenoh_ros2dds` the payload topic never arrived in either run: the bridge routes a topic once a local reader exists and the reader waited for the bridge, leaving only `Pending=1` in a log.

rosotacom_dev-2.5.dev77/docs/findings/a-shorter-group-does-not-lose-fewer-key-frames.md

@@ -0,0 +1,116 @@
+# A Shorter Group Does Not Lose Fewer Key Frames
+
+## Claim
+
+Shortening the group of pictures is the design lever the fragment law hands a
+video stream: fewer bytes in the largest message class, so a smaller fragment
+count, so a lower loss probability for the frame everything else depends on.
+The lever works on that probability and **not on the rate at which key frames
+go missing**, because the same change makes proportionally more of them.
+
+Three size distributions at one cadence, each standing in for a group length,
+under an unchanged link:
+
+| link | pattern | large msg | per second | loss of it | **lost per second** | offered |
+|---|---|---:|---:|---:|---:|---:|
+| p = 1 % | 4 small + 1 large | 38.1 kB | 2.00 | 29.52 % | **0.590** | 0.936 Mbit/s |
+| p = 1 % | 2 small + 1 large | 24.1 kB | 3.33 | 23.82 % | **0.794** | 0.915 Mbit/s |
+| p = 1 % | 1 small + 1 large | 14.1 kB | 5.00 | 12.45 % | **0.623** | 0.768 Mbit/s |
+| p = 2 % | 4 small + 1 large | 38.1 kB | 2.00 | 56.43 % | **1.127** | 0.935 Mbit/s |
+| p = 2 % | 2 small + 1 large | 24.1 kB | 3.34 | 37.95 % | **1.266** | 0.915 Mbit/s |
+| p = 2 % | 1 small + 1 large | 14.1 kB | 5.00 | 21.98 % | **1.099** | 0.768 Mbit/s |
+
+Read the two bold columns together. **The loss of the large message falls by
+more than half — 29.5 % to 12.5 % — while the rate at which large messages are
+lost does not move: 0.590, 0.794, 0.623 per second.** The lever's two factors
+cancel to within a few per cent, by construction rather than by accident: a
+message 2.7× smaller occupies 2.7× fewer packets, and there are 2.5× more of
+them.
+
+**The factor that does move the answer is the one the designer does not set.**
+Doubling the link's per-packet rate moves the same quantity by 1.8–1.9× in
+every pattern.
+
+The rule is therefore usable and its limit is exactly stated: it predicts
+*which* term changes and by how much, and it does not predict that the outcome
+improves. What a shorter group buys is the length of each hole, not their
+number — and, worth having, about 18 % less offered bandwidth for the same
+cadence.
+
+## Setup
+
+- Host pair / topology: one host, the packaged local benchmark rig — two
+  communication containers on their own Docker network, `tc` inside each
+  container's own netns (`--sudo-mode container`).
+- Session: `bench_1_1_capacity` from the packaged example project, one
+  `a->b:/bench_capacity` stream at 10 Hz with `--size-pattern`, CycloneDDS,
+  OTA QoS best_effort / KEEP_LAST depth 1. The patterns are
+  `4x5KB+1x38KB`, `2x5KB+1x24KB` and `1x5KB+1x14KB` — a large class against a
+  fixed small one, in the three ratios a group length of five, three and two
+  produces.
+- rosotacom SHA: measured at `7bb420c`; needs 2.5.dev74 or later for the fitted
+  delay table to arm.
+- Profiles: `tilt-iid` and `tilt-iid-worse` in
+  [`src/rosotacom/resources/examples/profiles.yaml`](../../src/rosotacom/resources/examples/profiles.yaml)
+  — the 2026-08-17 drive's own delay table at 25 ms ± 18 ms with 1 % and 2 %
+  independent per-packet loss. Nothing else differs between the two links.
+- Seed policy: netem `loss` is unseeded, `n=1` per cell; each cell offers about
+  2100 messages of which 420–1680 are of the large class, so the intervals on
+  the large-class figures are 3–5 points wide and on the per-second rates
+  correspondingly.
+
+## Evidence
+
+Evidence grade: per-message counts from the receiving peer's own RFC 0003
+transit records. Lost rows carry no payload and therefore no size, so the size
+classes are learnt from the delivered rows as the publisher's repeating pattern
+against the sequence number, and every row — delivered or lost — is then
+assigned by its position in that cycle. The window is taken in sequence numbers
+(first 300, last 50 dropped).
+
+The transfer function itself holds throughout, so the cancellation is not a
+breakdown of the model but a consequence of it: fitting one `(p, F)` over the
+mixed workload leaves residuals of 0.86–1.11 at p = 1 % and 0.99–1.03 at
+p = 2 %.
+
+Verification: manual, from a source checkout with Docker:
+
+```bash
+for prof in tilt-iid tilt-iid-worse; do
+  for patt in 4x5KB+1x38KB 2x5KB+1x24KB 1x5KB+1x14KB; do
+    rosotacom benchmark probe \
+      --project src/rosotacom/resources/examples/rosotacom.yaml \
+      --profile "$prof" --size-pattern "$patt" --rate-hz 10 --duration 240 \
+      --repeats 1 --sudo-mode container --no-plot
+  done
+done
+```
+
+Split each run's transit rows by size class and divide the large class's losses
+by the window in seconds. Within one link the three patterns must agree on that
+rate to within about a quarter; between the two links it must roughly double.
+
+## Status
+
+confirmed, 2026-08-31.
+
+## Publication notes
+
+This is the bench form of a field result that reads, without it, like an
+excuse. A campaign that shortened its group of pictures and then measured no
+improvement in the aggregate can attribute that to the link having been worse
+on the days that followed — which is true and unsatisfying, because it is
+unfalsifiable from one campaign. Here the link is held still by construction
+and the aggregate still does not move, so the cancellation is a property of the
+lever rather than of those evenings.
+
+Stated positively, it is the sentence that makes the design rule usable: the
+fragment law tells you what will change and by how much, and the thing it
+changes is not the thing an operator counts. For a video path the honest
+formulation is that a shorter group trades hole *length* against hole *count*
+at constant link, and pays for it in nothing — the offered bandwidth falls by
+about a fifth at the same cadence and quality target.
+
+The 2× row is the other half and belongs with it wherever the rule is offered
+as advice: the term the designer does not control moved the outcome by more
+than the term they do.

rosotacom_dev-2.5.dev77/docs/findings/neither-a-queue-nor-a-burst-tilts-the-size-law.md

@@ -0,0 +1,182 @@
+# Neither A Queue Nor A Burst Tilts The Size Law
+
+## Claim
+
+The field fit of `1-(1-p)^N` carries a residual that is systematic in size: the
+smallest class loses about half of what the fit predicts and the largest about
+2.9× more, a spread of roughly six across the ladder. Three candidate causes
+were emulated against the same measurement chain, and none of them produces it.
+
+| | 300 B | 1.1 kB | 3.1 kB | 6.1 kB | 12.1 kB | 38.1 kB |
+|---|---:|---:|---:|---:|---:|---:|
+| independent per packet | 1.05 % | 1.47 % | 3.47 % | 6.28 % | 11.23 % | 31.92 % |
+| a two-state process, same mean, 105 ms bad runs | 1.47 % | 1.76 % | 2.76 % | 5.09 % | 10.00 % | **21.98 %** |
+| + a token bucket a burst can fill (4 Mbit/s) | 1.14 % | 1.33 % | 2.57 % | 4.99 % | 13.43 % | 32.92 % |
+| a serialisation constraint (1.5 Mbit/s) | — | 1.62 % | — | — | 11.60 % | **97.54 %** |
+
+- **The chain is not the cause.** Under independent per-packet loss the fitted
+  law is reproduced with residuals of 1.20, 0.85, 1.01, 1.06, 0.97 and 1.00 —
+  flat over three decades of size. Whatever tilts the field, it is not the
+  wrapper, the join, or the way the fit is taken.
+- **Clustered loss is not the cause, and it bends the curve the other way.**
+  A two-state process at the same 1 % mean, with its chain rescaled per cell so
+  that a bad run is the same 105 ms of wall clock in every one of them, makes
+  the largest class lose **31 % less** than the independent law rather than
+  more: 21.98 % against 31.92 %, intervals 20.3–23.8 against 30.0–33.9. The
+  reason is the one that makes it a mechanism rather than a coincidence — a
+  38-packet message occupies about the length of one bad run, so its packets
+  share a fate and the message is lost once instead of thirty-eight times
+  independently. The field's residual has the opposite sign, so burstiness at
+  this scale does not explain it; it makes it harder to explain.
+- **A burst meeting a queue is not the cause.** A token bucket at 4 Mbit/s —
+  above the largest class's mean offered rate of 3.0 Mbit/s and far below its
+  instantaneous one, so a 38-packet message meets a queue and a 1-packet
+  message does not — changes nothing systematic: 1.09, 0.90, 0.74, 0.80, 1.20,
+  1.03 against the unqueued arm. The queue defers rather than drops, which is
+  the same statement as
+  [oversubscription queues, it does not lose](oversubscription-queues-not-losses.md),
+  seen per size.
+- **A serialisation constraint does not tilt it either — it removes the class.**
+  At 1.5 Mbit/s the one-packet message pays 10 % more than unshaped and the
+  13-packet message 3 % more, while the 38-packet message goes from 31.9 % to
+  **97.5 %**. That is not a slope, it is a threshold: below the constraint
+  nothing changes, above it the stream is gone.
+
+The last row is worth naming for what it is *shaped* like. A link whose
+transmission opportunities are a fraction of the wall clock presents from above
+as exactly this — a small probe crossing unharmed while a large burst does not
+— which is the signature an uplink moved onto a duty-cycled carrier would leave
+in an application's own numbers. This reproduces the signature; it does not
+demonstrate the mechanism, because a sustained rate cap and a duty cycle are
+not the same thing and this cell is oversubscribed on the mean as well as on
+the burst.
+
+## Setup
+
+- Host pair / topology: one host, the packaged local benchmark rig — two
+  communication containers on their own Docker network, `tc` inside each
+  container's own netns (`--sudo-mode container`).
+- Session: `bench_1_1_capacity` from the packaged example project, one
+  `a->b:/bench_capacity` stream at 10 Hz, CycloneDDS, OTA QoS best_effort /
+  KEEP_LAST depth 1.
+- rosotacom SHA: measured at `7bb420c`; needs 2.5.dev74 or later for the fitted
+  delay table to arm at all.
+- Profiles: `tilt-iid`, `tilt-iid-queued` and `tilt-iid-tight` in
+  [`src/rosotacom/resources/examples/profiles.yaml`](../../src/rosotacom/resources/examples/profiles.yaml).
+  All three carry the 2026-08-17 drive's own delay table at 25 ms ± 18 ms and
+  1 % independent per-packet loss; they differ only in the rate cap in front of
+  it — none, 4 Mbit/s, 1.5 Mbit/s.
+- Seed policy: netem `loss` is unseeded, `n=1` per cell, about 2100 messages
+  inside the measured window of each.
+
+## Evidence
+
+Evidence grade: per-message counts from the receiving peer's own RFC 0003
+transit records, windowed in sequence numbers (first 300 and last 50 dropped)
+so that the unshaped messages before the shaper is armed leave both the
+delivered and the lost class at once.
+
+The independent arm, with one `(p, F)` fitted by maximum likelihood over the
+whole ladder (`p = 0.870 %` per packet, `F = 872 B`):
+
+| wrapped | N | offered | lost | observed | 95 % CI | obs/pred |
+|---:|---:|---:|---:|---:|---:|---:|
+| 300 B | 1 | 2101 | 22 | 1.047 % | 0.69–1.58 | 1.20 |
+| 1.1 kB | 2 | 2102 | 31 | 1.475 % | 1.04–2.09 | 0.85 |
+| 3.1 kB | 4 | 2105 | 73 | 3.468 % | 2.77–4.34 | 1.01 |
+| 6.1 kB | 7 | 2103 | 132 | 6.277 % | 5.32–7.40 | 1.06 |
+| 12.1 kB | 14 | 2102 | 236 | 11.227 % | 9.95–12.65 | 0.97 |
+| 38.1 kB | 44 | 2102 | 671 | 31.922 % | 29.96–33.95 | 1.00 |
+
+The queued arm on the same ladder: 24, 28, 54, 105, 283 and 693 losses of about
+2100 offered, giving the second row of the claim table. The tight arm: 34 of
+2104, 244 of 2103, and 1826 of 1872.
+
+**Measuring the two-state arm took two attempts, and the reason is worth
+recording.** `netem`'s Gilbert–Elliott model advances its chain once per
+*packet*, not once per second, so a size ladder at a fixed message rate runs the
+chain at a different speed in every cell — 10 packets/s for a one-fragment
+message against 380 for a 38-fragment one. The cells then do not share a loss
+process. The arm measured that way duly reported a per-packet mean of 0.33 % in
+its smallest cell against a stationary 1.00 %, because that cell never sampled
+enough of the chain.
+
+Holding the offered *packet* rate constant instead was the obvious fix and it
+measured something else entirely: **the pipeline serves messages, not packets.**
+At 60 messages/s the one-packet cell lost 10.774 % under the bursty profile and
+10.626 % under the independent one, and at 380 messages/s 48.391 % and 48.430 %.
+Two different loss processes cannot agree to three digits; what both were
+reporting is the depth-1 OTA queue overwriting its own backlog before the link
+sees it. That is the depth-1 overwrite of bundled publications, which the
+operator harness's ledger records against a real drive, reached here by
+accident and against a synthetic one.
+
+What works is to leave the message rate at the field's 10 Hz and rescale the
+chain: multiplying `p` and `r` by `380/(N × 10)` in each cell leaves the
+stationary bad fraction at 3.92 % and the mean loss at 1 % while making the bad
+run 105 ms of wall clock everywhere, which is the drive's own figure. The six
+`bp-frag-*` profiles are that ladder.
+
+Verification: manual, from a source checkout with Docker:
+
+```bash
+for prof in tilt-iid tilt-iid-queued; do
+  for size in 200 1000 3000 6000 12000 38000; do
+    rosotacom benchmark probe \
+      --project src/rosotacom/resources/examples/rosotacom.yaml \
+      --profile "$prof" --size "$size" --rate-hz 10 --duration 240 \
+      --repeats 1 --sudo-mode container --no-plot
+  done
+done
+rosotacom benchmark probe --project src/rosotacom/resources/examples/rosotacom.yaml \
+  --profile tilt-iid-tight --size 38000 --rate-hz 10 --duration 240 \
+  --repeats 1 --sudo-mode container --no-plot
+```
+
+```bash
+for pair in 200:one 1000:two 3000:four 6000:six 12000:twelve 38000:thirtyeight; do
+  rosotacom benchmark probe \
+    --project src/rosotacom/resources/examples/rosotacom.yaml \
+    --profile "bp-frag-${pair##*:}" --size "${pair%%:*}" --rate-hz 10 \
+    --duration 240 --repeats 1 --sudo-mode container --no-plot
+done
+```
+
+The independent ladder must fit one `(p, F)` to within about ten per cent at
+every point; the queued ladder must not differ from it systematically; the
+rescaled two-state ladder must fall below it at the large end and not above; and
+the tight cell at 38 kB must be near-total loss while its 1 kB cell is not.
+
+## Status
+
+confirmed, 2026-08-31.
+
+## Publication notes
+
+This is a set of negative results and the negative is the point: a residual
+that survives three plausible explanations is worth more than one that has not
+been attacked. What it licenses is a narrower claim than "we do not know why" —
+the tilt is not an artefact of the accounting, it is not what a queue does to a
+burst, and it is not a serialisation limit, because a serialisation limit does
+not tilt a curve, it truncates it.
+
+The clustered arm is the one that changes the reading rather than confirming
+it. A reviewer's first guess for a size-systematic residual is burstiness, and
+burstiness at the drive's own time scale produces the **opposite sign**: it
+protects large messages, because a message long enough to span a bad run is lost
+once rather than in each of its packets. So the field's residual is not
+explained by the mechanism that the field's own two-layer model already
+contains, which is worth saying out loud.
+
+What is left to attack is correlation at a scale *shorter* than a large
+message's own time on the wire — short enough that a message straddles a loss
+event more often than a single packet does, without its packets sharing one.
+`netem` cannot express that at all: its chain advances per packet, so a run is
+always measured in packets rather than in milliseconds, and the two coincide
+only at one message size.
+
+The serialisation row deserves one sentence wherever a duty-cycled uplink is
+discussed. It shows that the observable signature of such a link — small
+messages unaffected, large messages gone, round-trip time unchanged — is
+reproducible from a rate constraint alone, which means the signature on its own
+does not identify the mechanism.

rosotacom_dev-2.5.dev77/docs/findings/the-size-law-belongs-to-the-transport-not-the-link.md

@@ -0,0 +1,140 @@
+# The Size Law Belongs To The Transport, Not To The Link
+
+## Claim
+
+On one emulated link, with the impairment held byte-identical and only the OTA
+middleware changed, message loss against message size takes three different
+shapes:
+
+| wrapped size | fragments | CycloneDDS | Fast DDS | zenoh native |
+|---:|---:|---:|---:|---:|
+| 300 B | 1 | 1.05 % | 1.05 % | 0.57 % |
+| 1.1 kB | 2 | 1.47 % | 1.14 % | 1.00 % |
+| 3.1 kB | 4 | 3.47 % | 3.80 % | 0.71 % |
+| 6.1 kB | 7 | 6.28 % | 4.81 % | 1.00 % |
+| 12.1 kB | 13 | 11.23 % | 8.74 % | **0.47 %** |
+| 38.1 kB | 38 | 31.92 % | 39.25 % | **61.69 %** |
+
+**CycloneDDS follows `1-(1-p)^N` and nothing else.** Fitting one per-packet
+rate and one effective packet size over the whole ladder leaves residuals of
+1.20, 0.85, 1.01, 1.06, 0.97 and 1.00 — flat across three decades of size and
+a factor of 44 in packet count, with no trend.
+
+**Fast DDS keeps the form and loses the fit.** It carries no RTPS fragment size
+at all, so the effective packet is the datagram; one `(p, F)` describes the
+ladder to within 0.71–1.14, and the misfit is systematic rather than noise
+(at 12 kB the Wilson interval is 7.61–10.02 % against a predicted 12.32 %).
+
+**zenoh does not obey the law.** Its loss is flat at 0.5–1.0 % from 300 B to
+12 kB — a 13-packet message loses what a 1-packet message loses, which is what
+a transport that repairs its own losses looks like — and then falls off a cliff
+to 61.7 % at 38 kB, where the offered rate leaves the regime the repair can
+carry.
+
+So the transfer function this project reports from the field is not a property
+of the link. It is what a **best-effort, fragmenting, unrepaired** transport
+does to a link, and the same link answers a repairing transport completely
+differently. Any design rule of the form *a stream's availability follows from
+its largest message class* carries that scope with it.
+
+## Setup
+
+- Host pair / topology: one host, the packaged local benchmark rig — two
+  communication containers on their own Docker network, `tc` inside each
+  container's own netns (`--sudo-mode container`), no host privileges.
+- Session: `bench_1_1_capacity` from the packaged example project, one
+  `a->b:/bench_capacity` stream at 10 Hz, OTA QoS best_effort / KEEP_LAST
+  depth 1, `--rmw` the only thing that differs between the three ladders.
+- rosotacom SHA: measured at `7bb420c`; re-runs record theirs. The runs need
+  2.5.dev74 or later, because before it a profile carrying a fitted delay table
+  could not be armed at all
+  ([a fitted delay table never reached the shaper](a-fitted-delay-table-never-reached-the-shaper.md)).
+- Profile: `tilt-iid` in
+  [`src/rosotacom/resources/examples/profiles.yaml`](../../src/rosotacom/resources/examples/profiles.yaml)
+  — the 2026-08-17 drive's own delay table at 25 ms ± 18 ms, with independent
+  per-packet loss raised to 1 % so that a 240 s run resolves the one-packet
+  cell. The *form* of the law is under test, and it does not depend on the
+  level.
+- Seed policy: netem `loss` is unseeded; each cell offers about 2100 messages
+  inside the measured window, so the Wilson intervals in the table below carry
+  the run-to-run spread. `n=1` per cell.
+
+## Evidence
+
+Evidence grade: per-message counts read from the **receiving** peer's own RFC
+0003 transit records, not from the runner's summary — a summary over rows from
+two observers is exactly what has been wrong in this harness before
+([sender rows](sender-rows-make-a-dead-link-look-loss-free.md)). The window is
+taken in sequence numbers (first 300, last 50 dropped), because a lost row
+carries no timestamp of its own and the shaper is armed after discovery.
+
+CycloneDDS, one `(p, F)` fitted by maximum likelihood over the whole ladder —
+`p = 0.870 %` per packet, `F = 1008–872 B` against a declared 1024 B fragment,
+the difference being the per-packet header the fragment size does not count:
+
+| wrapped | N | offered | lost | observed | 95 % CI | predicted | obs/pred |
+|---:|---:|---:|---:|---:|---:|---:|---:|
+| 300 B | 1 | 2101 | 22 | 1.047 % | 0.69–1.58 | 0.870 % | 1.20 |
+| 1.1 kB | 2 | 2102 | 31 | 1.475 % | 1.04–2.09 | 1.733 % | 0.85 |
+| 3.1 kB | 4 | 2105 | 73 | 3.468 % | 2.77–4.34 | 3.435 % | 1.01 |
+| 6.1 kB | 7 | 2103 | 132 | 6.277 % | 5.32–7.40 | 5.934 % | 1.06 |
+| 12.1 kB | 14 | 2102 | 236 | 11.227 % | 9.95–12.65 | 11.515 % | 0.97 |
+| 38.1 kB | 44 | 2102 | 671 | 31.922 % | 29.96–33.95 | 31.921 % | 1.00 |
+
+Fast DDS on the identical profile, `p = 1.306 %`, `F = 1224 B`: 0.80, 0.87,
+0.98, 0.76, **0.71**, **1.14**. The two ends of the ladder sit on opposite
+sides of the fit and their intervals exclude it, so a single effective packet
+size does not describe this stack.
+
+zenoh native on the identical profile: 0.571, 0.998, 0.712, 0.996, 0.475,
+**61.693 %**. No fit is reported because there is nothing of this family to
+fit — five points are flat and the sixth is a cliff.
+
+**The effective packet is smaller than the fragment, and by a measurable
+amount.** CycloneDDS's fitted `F` lands at 872–1008 B against a configured
+1024 B fragment, i.e. every packet carries roughly 15 % of itself in headers.
+That is the same quantity as the wire-to-payload ratio measured on the tunnel
+(1.13–1.17×), reached from the opposite direction: one from counting bytes on
+an interface, one from fitting a loss curve.
+
+Verification: manual, and it needs a source checkout with Docker:
+
+```bash
+for size in 200 1000 3000 6000 12000 38000; do
+  for rmw in cyclone fastdds zenoh; do
+    rosotacom benchmark probe \
+      --project src/rosotacom/resources/examples/rosotacom.yaml \
+      --profile tilt-iid --size "$size" --rate-hz 10 --duration 240 \
+      --repeats 1 --rmw "$rmw" --sudo-mode container --no-plot
+  done
+done
+```
+
+Read the per-size counts from `logs/b/status/events.jsonl` of each instance
+(`kind: transit`, `status` delivered or lost, `size_bytes` on the delivered
+rows) rather than from `result.json`. The CycloneDDS ladder must fit one
+`(p, F)` to within about ten per cent at every point; the zenoh ladder must be
+flat below 12 kB.
+
+## Status
+
+confirmed, 2026-08-31.
+
+## Publication notes
+
+The useful sentence is the scope one. `1-(1-p)^N` is textbook and the field fit
+of it is a real result, but reporting it without naming the transport invites
+the reading that it describes the *link*. It does not: on one link, at one
+impairment, three transports produce a clean power law, a distorted one, and no
+law at all. A paper that quotes a per-fragment rate should therefore quote the
+middleware and its QoS beside it, and a design rule derived from the law should
+say that it holds for a best-effort path that fragments and does not repair.
+
+Two smaller points travel with it. The **effective packet is not the fragment**:
+the fitted size falls about 15 % below the configured one, which is the header
+share, and it is the same 15 % an interface counter reports as wire-over-payload
+— so a fragment budget computed from the configured size understates exposure
+by that much. And zenoh's shape is the more interesting half of the comparison
+rather than a footnote: a transport that repairs turns a size-dependent
+availability problem into a capacity cliff, which is a different failure to
+design against, not a smaller one.

{rosotacom_dev-2.5.dev75 → rosotacom_dev-2.5.dev77}/src/rosotacom/resources/examples/profiles.yaml

@@ -169,3 +169,150 @@ profiles:
   # camera stream - a source-side case. Deliberately unshaped: play the case
   # bag with `--profile none` to see that the link is not the cause.
   field-20260817-case6-source-gap: {}
+
+  # ---- fragment-law tilt (B1): three loss processes, one delay model -------
+  # The field fit says message loss follows 1-(1-p)^ceil(size/F) and the field
+  # residual is systematic in size: the smallest class loses about half of what
+  # the fit predicts and the largest about 2.9x more. The campaign cannot
+  # separate an accounting artefact from a property of the link. These three
+  # can: the delay model is the drive's own table in all of them, and the only
+  # thing that moves is how the loss is distributed over packets.
+  #
+  # The rate is raised to ~1% per packet in every arm so a 240 s run at 10 Hz
+  # resolves the small end; the law's *form* is what is under test, and it does
+  # not depend on the level.
+
+  # A -- independent per packet. The law is exact by construction here, so this
+  # arm measures the measurement chain and nothing else.
+  tilt-iid:
+    uplink:
+      delay: 25ms
+      jitter: 18ms
+      distribution: field-20260817
+      distribution_file: dist/field-20260817.dist
+      loss: 1%
+    downlink: { delay: 7ms, jitter: 5ms }
+
+  # B -- the drive's own two-state process, scaled to the same mean. Clustering
+  # correlates the fragments of one message, so large messages should lose
+  # *less* than the independent law predicts -- the opposite sign to the field.
+  # p and r are the drive's, so the runs and gaps keep their measured lengths.
+  tilt-bursty:
+    uplink:
+      delay: 25ms
+      jitter: 18ms
+      distribution: field-20260817
+      distribution_file: dist/field-20260817.dist
+      loss_gemodel: { p: 0.102%, r: 2.5%, loss_bad: 13.5%, loss_good: 0.49% }
+    downlink: { delay: 7ms, jitter: 5ms }
+
+  # C -- A plus a queue a burst can fill. 4 Mbit/s is above the largest class's
+  # mean offered rate (38 kB at 10 Hz is ~3.0 Mbit/s of payload) and below its
+  # instantaneous one, so a 38-fragment message meets a queue and a 1-fragment
+  # message does not. That is the burst-tilt candidate, isolated.
+  tilt-iid-queued:
+    uplink:
+      rate: 4mbit
+      delay: 25ms
+      jitter: 18ms
+      distribution: field-20260817
+      distribution_file: dist/field-20260817.dist
+      loss: 1%
+    downlink: { delay: 7ms, jitter: 5ms }
+
+  # B2 -- the boundary of the design rule. Same delay model, two per-fragment
+  # rates: the one the rule is applied at, and one twice as bad. The rule says
+  # what the size distribution does; it does not say what the link will be.
+  tilt-iid-worse:
+    uplink:
+      delay: 25ms
+      jitter: 18ms
+      distribution: field-20260817
+      distribution_file: dist/field-20260817.dist
+      loss: 2%
+    downlink: { delay: 7ms, jitter: 5ms }
+
+  # B4 -- a serialisation constraint tight enough that a burst cannot fit in
+  # its own send interval, which is the shape an EN-DC uplink on a TDD carrier
+  # would present from above: a small probe survives what a 38-fragment burst
+  # does not.
+  tilt-iid-tight:
+    uplink:
+      rate: 1.5mbit
+      delay: 25ms
+      jitter: 18ms
+      distribution: field-20260817
+      distribution_file: dist/field-20260817.dist
+      loss: 1%
+    downlink: { delay: 7ms, jitter: 5ms }
+
+  # B1-prime -- the two-state process with its chain rescaled per cell.
+  #
+  # netem advances a Gilbert-Elliott chain once per *packet*, so the fitted p
+  # and r describe a process in packets and not in time. At the drive's own
+  # 380 fragments/s they mean a bad run of about 105 ms; on a size ladder at a
+  # fixed 10 Hz the packet rate is N x 10, so the same p and r stretch the bad
+  # run by 380/(N x 10) and every cell meets a different link. Forcing the
+  # packet rate constant instead was tried and measured the wrong thing: the
+  # pipeline serves messages, not packets, and at 60 messages/s the depth-1 OTA
+  # queue overwrites its own backlog -- both loss processes then reported
+  # 10.7 % to two digits.
+  #
+  # So the chain is rescaled instead: p and r are multiplied by 380/(N x 10) in
+  # each cell, which leaves the stationary bad fraction at 3.92 % and the mean
+  # loss at 1 % while making the bad run 105 ms of wall clock everywhere. The
+  # comparison against `tilt-iid` is then a comparison of one thing.
+
+  bp-frag-one:
+    uplink:
+      delay: 25ms
+      jitter: 18ms
+      distribution: field-20260817
+      distribution_file: dist/field-20260817.dist
+      loss_gemodel: { p: 3.876%, r: 95%, loss_bad: 13.5%, loss_good: 0.49% }
+    downlink: { delay: 7ms, jitter: 5ms }
+
+  bp-frag-two:
+    uplink:
+      delay: 25ms
+      jitter: 18ms
+      distribution: field-20260817
+      distribution_file: dist/field-20260817.dist
+      loss_gemodel: { p: 1.938%, r: 47.5%, loss_bad: 13.5%, loss_good: 0.49% }
+    downlink: { delay: 7ms, jitter: 5ms }
+
+  bp-frag-four:
+    uplink:
+      delay: 25ms
+      jitter: 18ms
+      distribution: field-20260817
+      distribution_file: dist/field-20260817.dist
+      loss_gemodel: { p: 0.969%, r: 23.75%, loss_bad: 13.5%, loss_good: 0.49% }
+    downlink: { delay: 7ms, jitter: 5ms }
+
+  bp-frag-six:
+    uplink:
+      delay: 25ms
+      jitter: 18ms
+      distribution: field-20260817
+      distribution_file: dist/field-20260817.dist
+      loss_gemodel: { p: 0.646%, r: 15.83%, loss_bad: 13.5%, loss_good: 0.49% }
+    downlink: { delay: 7ms, jitter: 5ms }
+
+  bp-frag-twelve:
+    uplink:
+      delay: 25ms
+      jitter: 18ms
+      distribution: field-20260817
+      distribution_file: dist/field-20260817.dist
+      loss_gemodel: { p: 0.323%, r: 7.917%, loss_bad: 13.5%, loss_good: 0.49% }
+    downlink: { delay: 7ms, jitter: 5ms }
+
+  bp-frag-thirtyeight:
+    uplink:
+      delay: 25ms
+      jitter: 18ms
+      distribution: field-20260817
+      distribution_file: dist/field-20260817.dist
+      loss_gemodel: { p: 0.102%, r: 2.5%, loss_bad: 13.5%, loss_good: 0.49% }
+    downlink: { delay: 7ms, jitter: 5ms }

{rosotacom_dev-2.5.dev75 → rosotacom_dev-2.5.dev77}/src/rosotacom_dev.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: rosotacom-dev
-Version: 2.5.dev75
+Version: 2.5.dev77
 Summary: Host CLI for ROS OTA communication sessions.
 Author: develNor
 License-Expression: BSD-3-Clause

{rosotacom_dev-2.5.dev75 → rosotacom_dev-2.5.dev77}/src/rosotacom_dev.egg-info/SOURCES.txt

@@ -63,6 +63,7 @@ docs/testing.md
 docs/work-items.md
 docs/findings/README.md
 docs/findings/a-fitted-delay-table-never-reached-the-shaper.md
+docs/findings/a-shorter-group-does-not-lose-fewer-key-frames.md
 docs/findings/a-transport-element-doubles-every-fastdds-sample.md
 docs/findings/cyclone-fragment-equal-to-max-drops-fragmented-samples.md
 docs/findings/cyclone-spdp-discovery-bursts.md
@@ -74,10 +75,12 @@ docs/findings/jitter-loss-bandwidth-latency.md
 docs/findings/keyframe-loss-kills-the-next-deltas.md
 docs/findings/lifespan-is-enforced-by-one-reader-and-not-the-other.md
 docs/findings/lighter-message-loss.md
+docs/findings/neither-a-queue-nor-a-burst-tilts-the-size-law.md
 docs/findings/one-domain-two-cyclone-configs-stop-interoperating.md
 docs/findings/oversubscription-queues-not-losses.md
 docs/findings/reorder-becomes-reader-loss.md
 docs/findings/sender-rows-make-a-dead-link-look-loss-free.md
+docs/findings/the-size-law-belongs-to-the-transport-not-the-link.md
 docs/findings/zero-loss-boundaries-18kb-20hz.md
 docs/findings/repro/decode_chain_repro.py
 docs/findings/repro/reorder_drop_repro.py

{rosotacom_dev-2.5.dev75 → rosotacom_dev-2.5.dev77}/src/rosotacom_dev.egg-info/scm_file_list.json

@@ -47,6 +47,7 @@
     "docs/ffmpeg-keyframes.md",
     "docs/findings/README.md",
     "docs/findings/a-fitted-delay-table-never-reached-the-shaper.md",
+    "docs/findings/a-shorter-group-does-not-lose-fewer-key-frames.md",
     "docs/findings/a-transport-element-doubles-every-fastdds-sample.md",
     "docs/findings/cyclone-fragment-equal-to-max-drops-fragmented-samples.md",
     "docs/findings/cyclone-spdp-discovery-bursts.md",
@@ -58,12 +59,14 @@
     "docs/findings/keyframe-loss-kills-the-next-deltas.md",
     "docs/findings/lifespan-is-enforced-by-one-reader-and-not-the-other.md",
     "docs/findings/lighter-message-loss.md",
+    "docs/findings/neither-a-queue-nor-a-burst-tilts-the-size-law.md",
     "docs/findings/one-domain-two-cyclone-configs-stop-interoperating.md",
     "docs/findings/oversubscription-queues-not-losses.md",
     "docs/findings/reorder-becomes-reader-loss.md",
     "docs/findings/repro/decode_chain_repro.py",
     "docs/findings/repro/reorder_drop_repro.py",
     "docs/findings/sender-rows-make-a-dead-link-look-loss-free.md",
+    "docs/findings/the-size-law-belongs-to-the-transport-not-the-link.md",
     "docs/findings/zero-loss-boundaries-18kb-20hz.md",
     "docs/forensics-report.md",
     "docs/geo-link-map.md",